Changes:
- Added `GCHandleProxy` to debug builds
- this is used to track all `GCHandle` `Alloc` and `Free` calls to ensure that all allocations are freed.
- added some unit tests to make sure this is actually enforced
- as a result, several object are now freed correctly
- Added `ISKReferenceCounted` and `ISKNonVirtualReferenceCounted` interfaces to represent the reference counting types used in the native library
- this helps with automatically de-referencing objects
- `SKAbstractManagedStream`, `SKAbstractManagedWStream` and `SKDrawable` have been re-written to use better delegates
- instead of passing each of the delegates as parameters, they are now a struct that is passed as a single object
- better for extensions (which there shouldn't be) and only a single static field on the type
- removed the usage of `Marshal.GetFunctionPointerForDelegate`, which should help out with WASM (see #876)
- the objects now only keep weak references, meaning that they can now be garbage collected
- instead of trying to resolve the instances with a dictionary, a delegate is used and passed as "user context"
- Moved some of the repetitive logic from the types into the base `SKObject` and `SKNativeObject`
- some logic is automatically executed if the concrete type is `ISKReferenceCounted` or `ISKNonVirtualReferenceCounted`
- with the more centralized logic and stricter patterns, better tests can be written to make sure all memory is freed correctly and timely
- `SKData`, `SKFontManager` and `SKTypeface` now correctly prevent disposal of the "static" instances
- `SKPaint` now references the `Shader`, `MaskFilter`, `ColorFilter`, `ImageFilter`, `Typeface` and `PathEffect` properties
- this prevents accidental collection, or non-collection when the object goes out of scope
- the `SKPath` iterators (`Iterator` and `RawIterator`) and op builder (`OpBuilder`) now correctly own and dispose their native objects
- `SKRegion` objects are now disposed on the native side
- `SKTypeface` construction from a `SKManagedStream` (via both `SKTypeface` and `SKFontManager`) now copy the contents of the .NET `Stream` into a native memory
- typeface construction requires multiple seeks (previously, the stream was copied only if it was non-seekable)
- it also requires "duplicating" the stream, which is not supported on .NET streams
- duplicates or forks of a stream means that each of the streams need to be read concurrently from different locations
- .NET streams can only have a single position
- Updated the NuGets used for the tests
- using the `Xunit.AssemblyFixture` and `Xunit.SkippableFact` NuGets instead of using the code directly
- removed the `Xunit.Categories` NuGet as it was preventing tests from running
This PR has a big set of changes that may be breaking due to bug fixes:
- The `SKAbstractManagedStream`, `SKAbstractManagedWStream` and `SKDrawable` no longer prevent the GC from collecting them. This means that if code no longer references them, they will be disposed.
- As far as I can tell, this should not be a problem for the streams as they are never kept around - they are just used for reading and writing and typically only need to live for as long as a single method, and then need to be disposed by the caller. The `SKTypeface` and `SKDocument` do keep it around for a bit, but then they also take ownership of the stream and keep a hard reference to the streams themselves. They will dispose the streams when they are disposed.
- `SKDrawable` is never kept around and is entirely a user-controlled object. If it goes out of scope, skia doesn't have a reference anyway.
- The `SKFontManager` and `SKTypeface` no longer use the managed streams (`SKManagedStream` or `Stream`) directly
- they make a copy.
- This is simply because skia streams can do things that are not possible for .NET - they can be read concurrently from different positions. If a `SKFileStream` or `SKMemoryStream` are passed, then the streams are not copied.
- Further optimizations can be made in the case of a `MemoryStream` or `byte[]` to not actually copy but use GC pinning to get a handle to the managed data and work with pointers. But this can be done later so that this PR can be merged and tested.